2 Introduction Perception is initiated by sensors. The focus is on vision (as opposed to hearing, sensing).How do we process info. provided by sensors?What can I infer about the world from this sequence of sensors?
3 Processing Sensor Data It has several uses:Manipulation.Navigation.Object Recognition.
5 Recognizing Patterns Definition. Pattern recognition is the “act of taking in rawdata and taking an action or category of the pattern”(Pattern Classification, Duda, Hart, and Stork).ComputerPattern RecognitionActionInput Signal
6 A Particular Example Fish packing plant Sort incoming fish on a belt according to two classes:Salmon orSea BassSteps:Preprocessing (segmentation)Feature extraction (measure features or properties)Classification (make final decision)
10 Decision Theory Most times we assume “symmetry” in the cost. (e.g., it is as bad to misclassify salmon as sea bass).That is not always the case:Case 1.Case 2.Sea bass can with piecesof salmonX Salmon can with pieces ofsea bass
11 Decision BoundaryWe will normally deal with several features at a time.An object will be represented as a feature vectorX =x1x2Our problem then is to separate the space of feature valuesinto a set of regions corresponding to the number of classes.The separating boundary is called the decision boundary.
13 Generalization The main goal of pattern classification is as follows: To generalize or suggest the class or action of objectsas yet unseen.Some complex decision boundaries are not good at generalization.Some simple boundaries are not good either.One must look for a tradeoff between performance and simplicityThis is at the core of statistical pattern recognition
16 Designing Pattern Recognition Systems Components in a system:Sensing devicesOften a transducer such as a camera or microphone (features: bandwidth,resolution, sensitivity, distortion, latency, etc.)Segmentation and groupingPatterns must be segmented (there may be overlapping).Feature extractionExtract features that simplify classification.Ideally: values similar in same category and different among categories.That is we need distinguishing features (invariant to transformations).
17 Designing Pattern Recognition Systems Components in a system:d) ClassificationUse feature vectors to assign an object to the right category.Ideally: determine probability of category membership for an object.Learn to handle noise.e) Post processingUse output of classifier to suggest an action.Classifier performance? Error rate?Minimize expected cost or “risk”.
19 Applying Pattern Recognition Systems Steps:Data CollectionUsually very time consumingFeature ChoicePrior knowledge is crucialModel ChoiceSwitch to new features, new classifierTrainingLearn from example patternsEvaluationAvoid overfitting
23 Translation of Coordinates Let (x,y,z) be a point in the image.Let (X,Y,Z) be a point in the scene.Then,-x/f = X/Z y/f = Y/zx = -fX/Z y = -fY/Z
24 Lenses Real cameras use a lens. More light comes in. (but not all can be in sharp focus).Scene points within certain range Zo canbe identified with sharp focus.This is called the depth of the field.
25 CCD Camera The image plane is subdivided into pixels. (typically 512x512).The signal is modeled by the variation inimage brightness over time.
29 Photometry of Image Formation The brightness of a pixel is proportional tothe amount of light directed toward the camera.Light reflected can be of two types:Diffusely reflected (penetrates below thesurface of the object and is re-emitted).b. Specularly reflected. Light is reflected fromthe outer surface of the object.
30 Photometry of Image Formation Most surfaces have a combination of diffuselyspecularly reflected light.This is the key to “modeling” incomputer graphics
34 Image Processing Typically there are problems: missing contours noise contours
35 Edge Detection Edges are curves in the image plane where there is a clear change of brightness.How do we detect edges?Consider the profile of image brightnessalong a 1-D cross-section perpendicularto an edge.
43 Summary We need to extract information from sensor data for activities such as manipulation,navigation, and object recognition.A signal is modeled by the variation inimage brightness over time.Light reflected can be diffusely reflected orspecularly reflected.Stereopsis is similar to motion parallax.What is machine learning?